Semi-supervised learning for image segmentation

In this thesis, a new approach to semi-supervised image segmentation is proposed, by implementing Unconstraint Least Square Importance Fitting (ULSIF) algorithm. ULSIF estimates importance value without determining probability distribution functions of the labeled and test class seperately. Different methods based on color or texture features are applied on semi-supervised image segmentation problems and similarities and differences between these methods have been explored. The performance of ULSIF based segmentation is compared with the state of art method one-class SVM in terms of segmentation accuracy and run time. According to the results, both color and texture based semi-supervised one-class ULSIF gave better results than one-class SVM does. Additionally, ULSIF algorithm is implemented as a two-class image segmentation method. In two class image segmentation methods, features of the foreground and background areas are used together during the learning stages. Results of supervised two-class ULSIF is compared with SVM and neural networks. According to the results, two-class SVM gave better results than two-class ULSIF and neural networks do.